Banks don’t have an AI problem. They have a memory problem.
McKinsey’s new CEO guide to AI in banking finds adoption racing ahead of enterprise value and names four gaps. The one that decides the rest is shared context: every agent rebuilding what the bank already knows. The fix is a governed memory, with meaning, permissions and corrections built in.
What happened
On 30 October McKinsey published Closing banking’s AI value gap: A CEO guide. The starting point is a paradox. Banks spend more on technology than the next four sectors combined, according to the firm’s Global Banking Annual Review, yet that spending still struggles to turn into productive growth.
The wider numbers come from McKinsey’s 2026 State of AI survey. Nearly nine in ten organizations now use AI regularly in at least one business function, and 44 percent are scaling it across the enterprise, up from 38 percent a year earlier. Only 37 percent report any positive enterprise-level earnings impact from AI, and just 6 percent qualify as high performers.
The authors name four gaps that keep value trapped inside individual deployments: value, intelligence, execution, and risk and governance. They then set out six decisions for the CEO, from a founder’s mindset to a common enterprise AI stack. Two lines stood out to me. Many deployed AI systems “perform the same on day 300 as they did on day one.” And the test of progress: is the tenth AI transformation materially better or cheaper because of what the first nine built and learned?
My view
The diagnosis is right, and the intelligence gap is the one I’d fix first, because the others depend on it.
The article puts it plainly. Terms such as customer, income, exposure and delinquency are defined differently across products, functions and systems. People reconcile those differences with judgment. AI systems can’t do that reliably at scale, so each initiative rebuilds its own context and business rules. The result is a bank with hundreds of AI implementations and almost no shared capital.
You can see it from the customer’s side. Picture one customer meeting four agents in a year:
Each failure in the top row looks like a model problem. It’s a meaning problem first. The onboarding agent doesn’t know income was verified. Servicing treats “customer” as the account holder. Collections uses gross income because nobody told it the bank’s affordability rule uses net.
What I’d add to the prescription
McKinsey calls the remedy enterprise memory: accumulated knowledge orchestrated into shared language. I agree, with three conditions that decide whether it holds up in a regulated bank.
- Meaning is owned and versioned. “Income” in the memory is a definition with an owner, an effective date and a version, not whatever the last project wrote down. Changing it is a release, with the consumers known in advance.
- Permission is checked when an agent reads. Shared memory must never widen access. The servicing agent and the collections agent can read the same customer record and still see different fields, for different purposes, under different retention rules. Sharing the memory isn’t the same as sharing the entitlements.
- Corrections flow back as reviewed changes. The article is right that feedback loops are missing. The cheapest loop already exists: the reviewer who fixes an agent’s output. Record that correction against the definition or context record it exposed, and the next deployment inherits the fix. Skip that step and day 300 looks like day one.
That last point is what turns the tenth-transformation test from a slogan into something you can measure. Count how many terms a new deployment reused instead of redefining, how many corrections were promoted into the shared model, and whether the time from idea to approved release fell.
Why it matters
For the industry. Most banks have funded agents one use case at a time. The meaning underneath them, which is the reusable part, sits in prompts and project documents that no one maintains. The article’s own example of reuse, underwriting capabilities carried into mortgages, small-business lending and onboarding, only works if the second team can trust the first team’s definitions. That trust has to be built deliberately, and before the agent count grows.
How it connects to our work. Context Before Query now has a page on enterprise memory, built around the drawing above: approved meaning, context records with source and expiry, and policy at read time. Chapter 12 of the book walks through the same banking journey. To try it in practice, Ontology Studio turns a team’s terms into reviewed, versioned definitions and releases, and Workflow Studio pins a release to a workflow and keeps every reviewer correction with its reason. Demo 1 shows the cost of skipping this: the same question, two answers, because two systems meant different things by one word.
The article ends on the CEO as integrator-in-chief. I’d give that CEO one concrete question for the next steering meeting: which ten definitions does every AI deployment in this bank depend on, and who owns each of them? If nobody can answer, the memory doesn’t exist yet.
Sources
- Closing banking’s AI value gap: A CEO guide, McKinsey & Company, 30 October 2026. Survey figures, the four gaps, enterprise memory and the tenth-transformation test are from this article.
Disclosure: I was previously McKinsey’s global firm chief data officer. The drawing in this post is my own.
Drafted with AI research assistance and reviewed by Richard White before publication. Linked sources are the record; check them for the full reporting.